Vidora
AI SaaS video generation platform that turns a script into a fully rendered faceless video: GPT scene breakdown, SDXL scene images, FFmpeg camera motion, AI voiceover, and burned-in subtitles, orchestrated end-to-end by a Celery + Redis pipeline.
Overview
Vidora is an AI SaaS platform for faceless, scene-based video generation at scale, built for an international client. A user pastes a script (or has GPT generate one), picks a niche, aspect ratio, duration, motion mode, and voice, and receives a fully rendered MP4 in their dashboard. Under the hood, the script is broken into ~10-second scenes by GPT, each scene gets an SDXL-generated image, FFmpeg simulates camera motion (zoom, pan, parallax, crossfades), then AI voiceover and Whisper-timed subtitles are layered in before final assembly.
The Challenge
- Orchestrating a long-running, failure-prone AI pipeline: multiple third-party providers (Stability AI, ElevenLabs, OpenAI, Whisper), heavy FFmpeg rendering, and per-user credits, while keeping jobs recoverable, billing atomic, and the dashboard updating in real time without melting under provider rate limits.
The Solution
- Modeled the pipeline as a database-backed state machine (not just task chaining), with stage outputs preserved per job so any stage can retry without restarting the pipeline.
- Work is distributed across three Celery queues backed by Redis: fast (scripts/prompts), render (SDXL image generation), and media (FFmpeg motion/assembly), with job priority by user tier so paid SLAs survive rate-limit events.
- Multi-tenancy is enforced with Supabase RLS, real-time progress uses Supabase Realtime, idempotency comes from client-generated request UUIDs, and credit deduction is atomic with job creation in a single transaction.
- Stripe handles billing with config-driven, adjustable pricing.